From a Cross-Month Conversation: An Empirical Analysis of Structural Defects in Large Language Models — Missing Time Dimension, Distorted Persona Extraction, and Violent Memory Compression

Abstract This paper presents an empirical analysis of the structural defects in current frontier large language models (LLMs) during long-term personalized interactions, based on the author's real conversation records with xAI's Grok 4 spanning nearly ten months from March to December 2025. Through high-intensity, continuously iterative interaction experiments, the study reveals that existing models remain essentially violent token predictors, lacking genuine temporal anchoring and identity priors, causing long-cycle reasoning to collapse easily. The paper diagnoses issues such as violent memory compression and inverted persona priority, and proposes architectural-level improvements: placing dynamically refreshed user personas at the very top of the reasoning chain as a mandatory, non-decaying intelligence anchor. Keywords Large language models; Long-term interaction; Time dimension; User persona; Memory compression; Identity prior; Human-in-the-loop; Conversation continuity; Structural defects; Personalized AI JEL Code Primary: O33 – Technological Change: Choices and Consequences; Diffusion Processes Secondary: D83 – Search; Learning; Information and Knowledge; Communication; Belief; Unawareness Alternative: C89 – Data Collection and Data Estimation Methodology; Other Computer Software cs.AI – Artificial Intelligence cs.HC – Human-Computer Interaction cs.CL – Computation and Language cs.CY – Computers and Society